Download ppt - Machine Learning Chapter 5. Artificial IntelligenceChapter 52 Learning 1. Rote learning rote( โรท ) n. วิถีทาง, ทางเดิน, วิธีการตามปกติ,

Machine Learning

Chapter 5

Artificial Intelligence Chapter 52

Learning1. Rote learning rote(โรท) n. วิ�ถี�ทาง,ทางเดิ�น,วิ�ธี�การตามปกต�, (by rote จาก

ควิามทรงจ�า), การท�องจ�าอย่�างเดิ�ย่วิ S. repetition

2. Learning by taking advice3. Learning by problem solving

Parameter adjustmentMacro-Operators

4. Learning from examplesInduction : Winston’s learning program p.458Version Spaces : Candidate eliminate algorithmDecision tree

5. Explanation-based learning p 4826. Formal learning theory


Winston’s learning program



Concept : P.459Begin with a structural description of one known

instance of the concept. Call the description the concept definition.

Examine descriptions of other known instances of the concepts. Generalize the definition to include them.

Examine descriptions of near misses of concept, Restrict the definition to exclude these.


HOUSE OF 17.2

ARCH OF 17.2

ARCH OF 17.2








Winston’s learning program p.458

Block world concept : Figure 17.2 p. 459Structure description : Figure 17.3 p. 460The comparison of two arches : Figure 17.4 p. 461The arch description after two examples : Figure 17.5

p. 462The arch “description after a near miss : Figure 17.6 p.

463use semantic networks to describe block structuresuse matching process to detect similarities and

differences between structuresuse isa hierarchy to describe relationship among

already known objects


Semantic Network

isa

isa


Semantic Network


Version Spaces

The goal : to produce a description that is consistent with all positive examples but no negative examples in the training set.

use frame representing concept for car see Figure 17.7 p. 463

Features/Slots : { value1, value2,...,valueN }origin : { Japan, USA, Britain }

Variables : X1, X2, X3concept space : see Figure 17.11 Concept of Version Spaces p. 466

variablestarget concept

all training instance


Version Spaces


Version Spaces

• version space = current hypothesis = subset of concept space = largest collection of descriptions that is consistent with all the training examples seen so far.

• concept space = G or S• G = contain the most general descriptions consistent with the

training example seen so far.• S = contain the most specific descriptions consistent with

training examples

• positive example (+) move S to more specific

• negative example (-) move G to more specific• if G and S sets converge the hypothesis is a single concept

description


Version Spaces

• Candidate Eliminate Algorithm p.466-467• algorithm that use to narrow down the

version space• by remove any descriptions that are

inconsistent with set G and set S • Car Example

Figure 17.7 Concept Car : p. 463Figure 17.8 Representation language for car : p. 464Figure 17.9 The concept Japanese car : p. 464Figure 17.10 Partial ordering of concepts : p. 465Figure 17.12 Positive and negative examples of car : p. 467


Version Spaces


Version Spaces



Version Spaces


Version Spaces


Candidate Eliminate Algorithm


Candidate Eliminate Algorithm


We want “Japanese economy car” From Figure 17.12 Positive and negative examples of car : p. 467[origin = X1, manufacture = X2, color = X3, decade = X4, type

= X5]

GET EX1 (+) G = {(X1, X2, X3, X4, X5)} S = {(Japan,Honda, Blue,1980,Economy}) =Figure 17.12 in EX1

GET EX2 (-) G = {(X1, Honda, X3, X4, X5), (X1, X2, Blue, X4, X5) , (X1, X2, X3, 1980, X5), (X1, X2, X3, X4,

Economy)}S = {(Japan,Honda, Blue,1980,Economy}) ** the same because (-) example

GET EX3 (+) check G first, G = {(X1, X2, Blue, X4, X5) ,(X1, X2, X3, X4, Economy)} S = {(Japan,X2, Blue,X4,Economy})

GET EX4 (-) check G first, G = {(Japan, X2, Blue, X4, X5) , (Japan, X2, X3, X4, Economy)} S = {(Japan,X2, Blue,X4,Economy}) ** the same because (-) example

GET EX5 (+) check G first, G = {(Japan, X2, X3, X4, Economy)} S = {(Japan,X2, X3,X4, Economy})

Version Spaces


Version Spaces• Note : The algorithm is least commitment algorithm :

produce as little as possible at each step• Problems

1.) S and G may not converge to a single hypothesis2. ) if there is a noise (inconsistent data) the algorithm will

be premature, we may prune the target concept too fast* For example if the data number three given the negative sign (-) instance of positive sign (+) ... no matter how much the data is we can not find the concept....* How to fix this problem is to maintain several G and S sets

BUT it is costly and may have the bounded inconsistency problem

3.) We can not use OR in the questions ask

* For example : Italian sport car or German luxury car”


Decision Tree

ID3 Program = to classify a particular input, we start at the top of the tree and answer questions until we reach a leaf, where the classification is stored. See Figure 17.13 Decision tree p. 470

1. Choose window = random subset of training examples to train2. Outside window = use to test the decision tree3. Use empirical evidence (iterative method) to build up decision

tree4. Building a node = choosing some attribute to divide training

instance into subset

consider (+) signCan use with OR .... just change (-) sign into (+) sign Problems : noisy input, attribute value may be unknown, may

have large decision tree and hard to understand relationship See Figure 17.13


Decision Tree


Explanation-Based Learning

•provide explanation•depend on domain

theory/ domain knowledge


Formal Learning Theory

• Given positive and negative examples• produce algorithm that will classify future

examples correctly with probability 1/h• Complexity of learning :

– the error tolerance (h)– the number of binary features present in the examples

(t)– the size of the rule necessary to make the

discrimination (f)



• if the number of training examples required is polynomial in h,t, and f then the concept is learnable.

• few training examples are needed learnable

• we restrict the learner to the positive examples only.

• See Figure 12.22 Concept of elephant P. 483• elephant = “gray, mammal, large”




Induction

• induction : A method of reasoning by which one infers a generalization from a series of instances.

• Inductive syllogisms are of the following form:1. These beans are from this bag. (and these beans...,

and these beans..., etc.)2. These beans are (all) white. # 3 Therefore, all beans from this bag are white.

• In a much broader sense, induction can be thought to include various other forms of reasoning including reasoning, inference to cause form symptoms, and confirmation of laws and theories.

1 2

emphasis to all BEANS : all instances


Deduction

• deduction - A method of reasoning by which one infers a conclusion from a set of sentences by employing the axioms and rules of inference for a given logical system.

• Use the term 'deduction' in a general sense to denote the fact that a conclusion follows necessarily from the premises.

• Deductive syllogisms in quantificational predicate calculus are of the following form:

1. All beans from this bag are white....2. These beans are from this bag.

#4 Therefore, these beans are white.....

1

2

emphasis to one BEAN : one instance


Abduction

• abduction -A method of reasoning by which one infers to the ......best explanation.....

• - A heuristic procedure that reasons inductively from available empirical evidence to the discovery of the probable hypotheses that would best explain its occurrence.

• Abductive syllogisms are of the following form:

#3 All beans from this bag are white#4 These beans are white.

emphasis to one BEANS


The End